Nvidia boss claims there’s a fix for dangerous AI ‘hallucinations’

According to Nvidia boss Jensen Huang, a central problem with the reliability of AI software can certainly be solved. One way is to let the software check the information. “Hallucinations are very solvable,” Huang claimed on the sidelines of the Nvidia developer conference GTC on Tuesday. AI hallucinations are cases of software simply inventing things.

The problem arises because programs such as the chatbot ChatGPT estimate word by word how a sentence could continue when formulating texts. This can sometimes cause them to give completely wrong answers, even though they have only been trained with the correct information.

Why is Nvidia fueling the AI ​​hype?

The semiconductor specialist Nvidia plays a key role in the current boom in artificial intelligence with its chip systems and benefits enormously from this economically.

At the GTC conference, Huang presented the next generation of the computing platform called Blackwell, on which AI applications will run more efficiently.

Huang was convinced that artificial intelligence would largely relieve humans of the burden of writing software code. For a while it was said that everyone should learn to program. “I think that’s wrong,” said the Nvidia boss. It wasn’t people’s job to learn a programming language. Instead, computers should be useful even if you don’t know a programming language.

However, you will have to learn to get the desired result from AI software using properly formulated requests – so-called ‘prompts’. But it is fundamentally similar when it comes to people’s ability to interact with each other: “When my wife talks to me, it happens quickly too – and it works perfectly.”

Huang thinks it is possible that within a few years, artificial intelligence will surpass human capabilities in several areas – for example in mathematics and bar exams or in understanding texts.

This is also what he relies on in the current debate about whether software will reach the level of so-called general artificial intelligence – which usually means that it can generalize and is superior to humans. However, there is no definition that everyone agrees on, he emphasized.

Critical experts see this differently: a technical breakthrough towards general artificial intelligence (AGI) is not foreseeable.

(rbu/sda/awp/dpa)